Qwen3.8 27B

Qwen/Qwen3.8-27B

27BDenseText generationImage → textText inputImage input

Qwen · 22 recorded configurations

Measured setup

Hardware

2× RTX 3090 · GPU-resident
1× RTX 3090 · GPU-residentalso Runtime · Quant
Placement unverifiedalso Runtime · Quant

Runtime and tool

Quantisation

MTP draft depth

Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.

PP · left axisTG · right axis

View exact measurements

Exact context depth and measured fresh prefill, bulk prefix, and decode rates for Qwen3.8 27B Uncensored OrcaRouter
Context depth (tokens)	Fresh prefill PP (tok/s)	TG (tok/s)
0	1,476.45	48.28
2,048	1,430.08	47.95
4,096	1,404.56	47.41
8,192	1,371.82	46.50
16,384	1,297.15	45.14
32,768	1,175.11	42.42
65,536	992.21	38.09
131,072	748.20	31.42
262,144	484.91	22.95

Concurrent-stream measurements

Concurrent-stream throughput at the recorded existing-prefix depth for Qwen3.8 27B Uncensored OrcaRouter
Streams	Aggregate PP (tok/s)	Aggregate TG (tok/s)	TG / stream (tok/s)	Aggregate all (tok/s)
1	1,476.45	48.28	—	—
1	1,376.93	48.20	48.20	135.14
2	1,495.30	84.98	42.49	228.92
4	1,539.39	121.48	30.37	314.76

Technical record and provenance

Run details

Source run ID: rtx3090x2-qwen3.8-27b-orcarouter-q6k-2026-08
Measured date: 2026-08-23
Hardware: 2× RTX 3090
Runtime: llamaCpp · llama-bench fc6545d (build 50)
Weight placement: gpu-resident
Split ratio: 1:1
KV-cache precision: Q8_0
Flash Attention: Enabled
Batch size: 2,048
Microbatch size: 2,048
Prompt tokens: 512
Generated tokens: 256
Speculative decoding: None (standard decoding)
Measured repetitions: 3
Notes: Q6_K Uncensored-OrcaRouter finetune, batch 2048/2048, all layers on GPU, tensor split 1:1. Concurrency rows from llama-batched-bench (aggregate rates), first single-stream pass discarded as warm-up. Values transcribed from the preserved raw llama-bench / llama-batched-bench JSON output. Model is not in the calculator catalog, so this record is archival until it is added.
Source article: /benchmarks/
Raw log path: research/benchmarks/logs/rtx3090x2/2026-08-23-qwen3.8-27b/

Compare measured entries